Currently when setting up an IMR around the kernel .text area we lock that IMR, preventing further modification. While superficially this appears to be the right thing to do, in fact this doesn't account for a legitimate change in the memory map such as when running through kexec. In such a scenario a second kernel can have a different size and location to it's predecessor and can view some of the memory occupied by its predecessor as legitimately usable RAM. This RAM can then be allocated to DMA agents within the system and trigger an IMR violation.
The solution to this situation is to keep the kernel .text section IMR lock bit false. This means that a subsequent kernel will boot and can tear-down an existing IMR, while still setting up an IMR around its own .text section. Signed-off-by: Bryan O'Donoghue <pure.lo...@nexus-software.ie> Reported-by: Andriy Shevchenko <andriy.shevche...@intel.com> --- arch/x86/platform/intel-quark/imr.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/arch/x86/platform/intel-quark/imr.c b/arch/x86/platform/intel-quark/imr.c index c1bdafa..1c43639 100644 --- a/arch/x86/platform/intel-quark/imr.c +++ b/arch/x86/platform/intel-quark/imr.c @@ -591,7 +591,7 @@ static void __init imr_fixup_memmap(struct imr_device *idev) * from the beginning of the .text secton to the end of the * .rodata section as one physically contiguous block. */ - ret = imr_add_range(base, size, IMR_CPU, IMR_CPU, true); + ret = imr_add_range(base, size, IMR_CPU, IMR_CPU, false); if (ret < 0) { pr_err("unable to setup IMR for kernel: (%p - %p)\n", &_text, &__end_rodata); -- 2.5.0